if you're running AI-powered outbound at any kind of scale, you've probably defaulted to MCP because it's the easiest path. i've seen it happen over and over with the brands and creator teams i work with. it makes sense - the tools are getting better fast. but when you actually look at the numbers, that choice can quietly bleed your budget dry.
Here's what nobody tells you upfront: each MCP tool loads its entire schema into context before your task even starts. that eats up nearly three-quarters of your available context before you've processed a single useful token. A basic enrichment job? over 40,000 tokens through MCP. the same job via CLI? just over 1,300. That's a 30x difference per run. now multiply that by a few thousand records a day. It stops being pocket change fast.
reliability is the other hidden killer. MCP sits around 70% reliability in my experience, CLI hits 100%. At low volumes you can shrug it off. At scale that 30% gap means hundreds of silent failures. prospects don't get enriched, sequences don't fire, and you only find out when something breaks downstream. then you're digging through logs trying to figure out what happened.
then there's chaining. MCP doesn't support native chaining - every step means a round trip. signal in, account enrichment, copywriting, push to sequence? that's four separate calls. with CLI you do it all in one pass. And because you can run CLI headlessly, you throw ten parallel workers at it while a single MCP call is still warming up. Set a cron job for 6 AM and your team wakes up to a fully enriched list, no manual kicking off required.
We rebuilt our whole stack on Claude Code with CLI. Signal comes in, Claude enriches the account, scores it against ICP, writes tailored copy for each vertical, and pushes it to sequence. zero human touch. it runs while everyone focuses on actual relationship-building.
look, MCP still has its place - deep research on a single account, browser automation, stuff that needs complex reasoning. but if you're leaning on it for enrichment and personalisation at scale, you're probably burning 30x more than you need to. At a certain point that stops being a tool preference and becomes an expensive habit.
